Skip to main content

Full text of "Unusual Dynamical Scaling in the Spatial Distribution of Persistent Sites in 1D Potts Models"

See other formats

Unusual Dynamical Scaling in the Spatial Distribution of 
Persistent Sites in ID Potts Models 




A. J. Bray and S. J. O'Donoghue 
Department of Physics and Astronomy, The University, Manchester, M13 9PL, United Kingdom 

(February 1, 2008) 

The distribution, n(k,t), of the interval sizes, k, between clusters of persistent sites in the 
dynamical evolution of the one-dimensional g-state Potts model is studied using a combination of 
numerical simulations, scaling arguments, and exact analysis. It is shown to have the scaling form 
n(k, t) — t 
the fraction 

the scaling length t e for the interval-size distribution is larger than the coarsening length scale t 1 ^ 2 
that characterizes spatial correlations of the Potts variables 

2z f(k/t z ), with 2 = max(l/2,#), where 9(q) is the persistence exponent which describes 
P(t) ~ t~ e , of sites which have not changed their state up to time t. When 6 > 1/2, 
















The discovery of persistence has recently generated 
considerable interest in understanding the statistics of 
first passage problems in spatially extended nonequi- 
librium systems, both theoretically and experimentally. 
The definition of persistence is as follows. Let <j>(x,t) 
be a stochastic variable fluctuating in space and time ac- 
cording to some dynamics. The persistence probability is 
simply the probability P(t) that at a fixed point in space, 
the quantity sgn [<f>(x, t) — {<j){x, t))] does not change up to 
time t. In many systems of physical interest a power law 
decay, P{t) ~ t~ e , is observed, where 9 is the persistence 
exponent and is, in general, nontrivial. The nontriviality 
of 9 emerges as a consequence of the coupling of the field 
<p(x, t) to its neighbours, since such coupling implies that 
the stochastic process at a fixed point in space and time 
is non-Markovian. 

Persistence phenomena have been widely studied in re- 
cent years [Qhi2| • Theoretical and computational studies 
include spin systems in one |^,|3| and higher dimen- 
sions, diffusion fields ||, fluctuating interfaces ||, phase- 
ordering dynamics , and reaction-diffusion systems Q . 
Experimental studies include the coarsening dynamics of 
breath figures H, soap froths |l0| , and twisted nematic 
liquid crystals [|Q| . Persistence in nonequilibrium critical 
phenomena has also been studied in the context of the 
global order parameter, M(t), (e.g. the total magnetiza- 
tion of a ferromagnet), regarded as a stochastic process 

In the present work we consider spatially extended sys- 
tems with a nonequilibrium field 4>(x, t), which takes dis- 
crete values, at each lattice site x. The field then evolves 
in time t through interactions with its neighbours. The 
persistence probability at time t is defined as the fraction 
of sites in which the stochastic field <p(x, t) did not change 
its value in the time interval [0, t\. The physical interpre- 
tation of <j)(x, t) could be for, for example, the coarsening 
spin field in the Ising model after being quenched to low 
temperature from an initial high temperature, the sign of 
a diffusion field starting from a random initial configura- 
tion or the sign of the height, relative to the mean height, 
of a fluctuating interface. As the stochastic field evolves 

in time, such systems develop regions of persistent and 
nonpersistent sites. Since the number of persistent sites 
decays with time according to P(t) ~ t , the persistent 
clusters shrink in size and number with a corresponding 
growth in the size of the nonpersistent regions. 

Recently, Manoj and Ray (MR) |l3| have studied such 
spatially extended systems in one dimension (ID) within 
the context of the A + A — > reaction diffusion model, 
which is equivalent to the ID Ising model. They found 
that the length scale which characterises the interval 
sizes between persistent clusters apparently has a differ- 
ent time-dependence from the length scale characterising 
the walker separations (or spin correlations, in the Ising 
representation), and futhermore seems to depend on the 
initial walker density. If true, this result would be sur- 
prising. Naively, one would expect the initial walker den- 
sity to be irrelevant to the asymptotic dynamics, and the 
coarsening length scale, set by the mean distance between 
walkers, to be the relevant length scale for all spatial cor- 
relations. We shall show that the former expectation is 
correct, but the latter, in general, is not. 

In this paper we generalise, and reinterpret, this study 
in the context of the g-state Potts model, which has q 
distinct but equivalent ordered phases. The A + A — ► 
model corresponds to the q = 2 Potts model, i.e. the 
Ising model, with the walkers identified as the domain 
walls between up and down Ising spins. In the g-state 
Potts model the random walkers represent domain walls 
between Potts states. At each time step, every walker 
hops randomly left or right. Persistent Potts sites are 
those at which the Potts state has never changed, i.e. 
sites which have never been crossed by a walker. If two 
walkers occupy the same site they either coalesce, with 
probability (g — 2)/(g — 1), or annihilate, with proba- 
bility l/(q — 1), these numbers being the probabilities 
that the states on the farther sides of the walkers are the 
same (annihilation) or different (coalescence). The per- 
sistence probability (the fraction of sites that have never 
been jumped over by a walker) decays as a power of time, 
with a g-dependent exponent, P(t) ~ t~ 9 ^ q \ 

Among the questions which emerge naturally in such 
a study are (i) What is the dominant length scale in 
the problem, as far as the persistent structures are con- 


cerned? (ii) What is the nature of the spatial distribution 
of the persistent sites? (iii) What is the average size of a 
persistent cluster? 

An important point to make at the outset concerns 
the different length scales associated with the walkers 
and with the persistent sites. The walker density decays 
as i -1 / 2 for any q, so the mean distance between walkers 
grows as L w (t) ~ t 1 / 2 . The density of persistent sites 
decays as t~ e , so the mean distance between these sites 
grows as L p (t) ~ t 6 . This simple observation immedi- 
ately suggests that the spatial structure of the persistent 
sites for 8 > 1/2, where L p is the larger length scale, will 
be very different from when 8 < 1/2, where L w is the 
larger length. This is precisely what we find: the charac- 
teristic length scale, Li n t, controlling the distribution of 
the intervals between clusters of persistent sites, is given 
by Li n t — max (L w , L p ) . For the Ising case studied ear- 
lier Q, one has § 9 = 3/8, which is 'close' to 1/2. We 
believe the proximity of these two exponents is responsi- 
ble for the apparently non-universal behavior in the Ising 
system reported in MR. 

Derrida et al. |3| have obtained an exact expression for 
9 for ID Potts models with arbitrary q: 

9(q) = 

8 7T 2 

5 — 7= — 



The value of q corresponding to 9 = 1/2 is q c = 2/[l + 
V2cos(V57r/4)] = 2.70528.., so 9{q) > 1/2 for all in- 
teger q > 3. Note that the probabilistic algorithm for 
implementing the Potts model through the annihilation 
or coalescence of random walkers (domain walls) allows 
<j to be a real number, q > 2, while an equivalent Ising 
spin representation (see section IV) of the Potts persis- 
tence problem even allows 1 < q < 2! By this means 
means we can explore a range of 8 above and below 1/2. 

Our main result is that the scale length controlling the 
distribution of interval sizes between persistent clusters is 
given by t 1/2 when 9 < 1/2, but by t e when 9> 1/2. We 
find no evidence for any dependence of the asymptotic 
scaling distribution on the initial walker density (other 
than through nonuniversal amplitudes). 

This paper is organised in the following manner. In 
section II we outline a scaling phenomenology within 
which the above questions can be addressed. In section 
III we give exact results for the q = oo Potts model. Fi- 
nally, in section IV, we present extensive numerical sim- 
ulations for the g-state Potts model which confirm our 
predictions based on general scaling arguments and the 
exact large-g results. Section V concludes with a discus- 
sion and summary of the results. 


The basic scaling phenomenology was introduced by 
Manoj and Ray, although we will adopt a slightly differ- 
ent notation. We are interested in the distribution of the 

sizes of the intervals between persistent sites. We define 
an interval size, k, as the number of non-persistent sites 
between two consecutive persistent sites. If the interval 
size is zero, the sites belong to the same cluster of per- 
sistent sites, and we henceforth consider only intervals of 
non-zero size. We define n(k,t) to be the number of in- 
tervals (per site) of size k at time t. A natural dynamical 
scaling assumption is 

n(k,t)=t- 2z f(k/t z ) 


where L{t) = t z is the 'characteristic' length scale at time 
t. (A dynamical exponent z is conventionally defined via 
L(t) ~ i 1 / 2 , rather than t z . Here we are following the 
notation of MR) . 

The rationale for the scaling form (||) is as follows. The 
number of non-persistent sites (per site) is given by 

Q(t) = ^fcn(ft,t) • 



But since the number of persistent sites (per site), P{t), 
decays to zero as P(t) ~ t~ s , and P(t) + Q(t) = 1, it 
follows that Q(t) — > 1 for t — > oo. Converting the sum 
([|) to an integral, with lower limit zero (valid as t — > oo 
provided the integral converges), we see that the prefac- 
tor t~ 2z in (||) is precisely what is needed to satisfy the 
required condition Q(t) — » const for t — > oo. 

Consider next the number, N c (t), of persistent clus- 
ters per site. Since the number of clusters is equal to the 
number of intervals, N c is given by 



Converting the sum to an integral (with lower limit zero) , 
using (||) and assuming the integral converges, gives 
N c ~ t~ z . Thus the mean distance between persistent 
clusters, i.e. the mean interval size, increases as t z . 

Is this reasonable? Let us recall that there are two 
length scales in the system, L w ~ t 1 / 2 , the mean dis- 
tance between walkers, and L p ~ t e , the mean distance 
between persistent sites. To make further progress we 
make the following two assumptions, both of which are 
confirmed by our numerical studies, by exact results for 
q = oo, and by heuristic arguments to be expounded be- 
low: (i) The mean cluster size tends to a constant for 
t — > oo, and (ii) The length scale Li nt = t z that controls 
the interval size distribution is the larger of L w and L pi 
i.e. z — max (1/2, 9). 

From assumption (i) we deduce that N c ~ t~ e , i.e. 
z = 9, which is consistent with assumption (ii) provided 
9 > 1/2. What if 8 < 1/2? Then we still require 
N c - t- e , but now z = 1/2, so the result N c - t~ z , 
derived from (|J) breaks down. Going back to (Q), we 
infer that this breakdown requires that the conversion of 
the sum to an integral be invalid - the sum must have 


its dominant contribution from k of order unity, rather 
than k of order t z . This in turn requires that the scaling 
function f(x) in (||) have a singular small- a; limit of the 
form 111 


with 1 < t < 2. Using this form in (|J), with z = 1/2, 
and inserting the result into (|]), gives iV c ~ t _ ( 1_T / 2 ). 
Comparing this with N c ~ t~ e fixes the value of t: 


< 1/2 


We first present a heuristic argument for the result 
z = max (1/2,6*). Consider first the case 8 > 1/2. In 
this regime, the distance between clusters is much greater 
that the distance between walkers, i.e. there are, on aver- 
age, many (~ < 0-1 / 2 ) random walkers between each con- 
secutive pair of persistent clusters, as illustrated in Figure 
|l|. However, walkers separated by distances large com- 
pared to i 1 / 2 are essentially uncorrelated, because any 
correlations are mediated by the random walkers, and 
correlations between walkers decay on the length scale 
i 1 / 2 . Therefore one expects the intervals between clus- 
ters, measured on the scale t , to be independent random 
variables, and the interval size distribution to scale with 
this length. This argument also suggests that the loca- 
tions of the clusters as a function of position on the lattice 
are described by a Poisson process, i.e. the interval size 
distribution is an exponential, n(k,t) = (k) exp(— k/(k)), 
with (As) ~ t e , for 6 > 1/2. These expectations are borne 
out by our numerical studies. 





<XX-a@ xx xskx©: x 

1 O O© <0) OOCIJ: 

FIG. 1. Snapshots showing the relative densities of persis- 
tent sites (0) and random walkers (x) for various q, at time 
t — 10 4 . The Ising representation was used to generate the 
q — 3/2 data, with the Potts representation used for the other 
q- values. 

For 8 < 1/2, the opposite is true: there are many clus- 
ters between each pair of walkers (see the q = 3/2 snap- 
shot in Figure Q). The largest length scale is set by the 
walker spacing, L w ~ i 1 / 2 , and the interval size distribu- 
tion will scale with this length. On smaller scales, there 

will be residual structure in the cluster distribution left 
from an earlier epoch when the number of walkers was 
larger. Indeed, on these smaller scales the persistent sites 
form a fractal set |l^-[l^]. Consider two sites separated 
by a distance r. The probability, P2(r,t), that both are 
persistent has the scaling form, 

P 2 (r,t) = t- 2e F(r/t^) , 



where F(x) — > const for x — > oo, since for r 3> t 1 / 2 the 
sites are uncorrelated. The residual structure remain- 
ing on scales r <C t 1 ^ 2 , however, means that if the first 
site is persistent, which occurs with probability of order 
t~ e , the probability that the second one is also persis- 
tent depends only on r. This implies F(x) ~ x~ for 
x — > 0. The number of persistent sites within a distance 
R of a given persistent site can therefore be estimated as 
dr r~ 2e ~ R d f , where 

1 - 26> 


is the fractal dimension of the persistent set. 

Clearly this result only makes sense for < 1/2, since 
df cannot be negative. This suggests d/ = 0for6 l >l/2, 
corresponding to point-like objects, i.e. isolated finite 
clusters. For q — oo, for example, where 9 — 1, we find 
that there is typically one cluster of persistent sites, with 
a fixed mean size, per scale length L p ~ t. 

We now present a heuristic argument in support of our 
claim that the mean cluster size, (I), approaches a con- 
stant as t — * oo for all values of 8. The initial steps of the 
argument follow the approach of MR. The total number 
of clusters per site can be written as 

N c (t) = P(t) - P w (t), 


where P w (t) is the fraction of walker sites which have 
never been visited by a walker, i.e. the persistence prob- 
ability for walker sites. Such sites form 'spacers' between 
each pair of persistent Potts states in a cluster, there be- 
ing exactly one more persistent Potts site than persistent 
walker site in each cluster. The mean cluster size is 


N c (t) 

P w {t) 



We expect, on universality grounds, that the exponents 
describing the decay of P(t) and P w (t) should be the 
same (i.e. 9), but the (nonuniversal) amplitudes will be 
different, i.e. P(t) -> At~ e , while P w {t) -> A w t~ e ', with 
A w < A. Inserting these forms in (10) gives the limiting 
value of the mean cluster size as (if^, = (1 — A w /A)^ 1 . 
This number is nonuniversal and is determined by the 
initial distribution of cluster sizes. 



Many of the general features of the dynamics for 8 > 
1/2 are exemplified by the q — > oo limit, for which the 
walker dynamics reduces to A + A — > A, i.e. the walkers 
always aggregate on contact. At t = all sites are per- 
sistent. The random walkers initially present divide the 
sites into clusters of persistent sites. Clearly no cluster 
can increase in size. Consider a cluster of initial size Iq. 
We first calculate the probability density pi (l,t), that 
the cluster survives to time t and has size I. The key 
point is that, for A + A — ► A dynamics, we need only 
consider the two walkers on either side of the initial clus- 
ter, since subsequent coalescence processes do not affect 
the random walk dynamics of these two walkers. The 
cluster and the two walkers can therefore be treated in 

For simplicity we begin by treating the continuum limit 
in which the walkers are regarded as continuous-time ran- 
dom walkers on a continuous space. This should be cor- 
rect in the limit Iq ^> 1 , an assertion we can subsequently 
test. Let the ends of the initial cluster (and the two 
walkers) be located at x = and x = Iq. Each walker 
obeys an equation of the form dx/dt — £(i), where £(t) 
is a Gaussian white noise with mean zero and correlator 
(£(*)£(*')) = 2S(t-t'). We first write down the probabil- 
ity distribution P r (x r ,t) of the rightmost excursion, x r , 
up to time i, of the left walker. An elementary calcula- 
tion gives 

P r (x r , £) 



ex P _ 77 


For the cluster to survive, we require x r < Iq, so in 
the limit t Iq the exponential factor in (11) can 
be dropped. The probability distribution, Pi(xi,t), of 
the leftmost excursion of the right walker (from its ini- 
tial position) is given by a similar expression. Clearly 
X\ Xy IS the residual size of the cluster at time 
t, where I < means the cluster has disappeared. The 
probability density pi (l,t) that the cluster has survived 
and has size I is therefore given, for t 3> Zq, by 

PloKht) = I I 0[lo - I - Xl - x r ) 


JO V TTt Jo V TTt 

(lo - I) 


The probability, p sur v(lo, t), that the cluster survives till 
time t is given by 

P (/.,/)= t 'di%-£ = ^ 



Immediately we see that the mean interval between 
surviving clusters grows as t (— t e , since 0(q — oo) = 1), 
which is much larger than the mean interval between 
walkers, which grows as t 1 ^ 2 . 

The mean length of surviving clusters, of given initial 
size Iq, in the t — > oo limit is 

(0°° — ]2 




If the initial clusters have a distribution of sizes, a 
straightforward calculation gives (l)^ = (/q)o/3(/q)o, 
where (. . .)o indicates an average over this distribution, 
while the fraction of clusters which survive is (Zq)o/27t£. 

The above calculations demonstrate that the mean 
cluster size approaches a constant at late times, and that 
the scale length for the sizes of the intervals between clus- 
ters grows as t, and not as the naive scaling length t 1 ! 2 
associated with the underlying domain wall coarsening. 
In fact we can calculate the interval size distribution, 
n(k,t), explicitly for this model. First recall that the 
surviving clusters are separated, at late times, by many 
(~ t 1 / 2 ) walkers. The fate of a given cluster depends only 
on the nearest walker on either side. The motion of these 
walkers is uncorrelated with that of the walkers bordering 
other clusters, because the correlation length for walkers 
grows only as t x l 2 . Therefore we can assume, in the scal- 
ing limit k — > oo, t — > oo, with k/t fixed, that clusters 
survive independently with probability l^jlTrt (where we 
have specialized to initial clusters of fixed size Iq). The 
probability distribution of the interval size k between 
neighboring clusters is therefore exponential, with mean 
(k) = 2-rrt/ll, while n(k,t) = (k)- 2 exp(-fc/(fc», which 
has the scaling form (|2|) with z = 1. This exponential 
form is in excellent agreement with simulation data pre- 
sented in section IV. 

Since the initial clusters used in the q — oo simula- 
tions are quite small (2 or 4), the continuum limit is not 
expected to be quantitatively correct. To conclude this 
section, therefore, we quote the asymptotic mean cluster 
size for clusters of arbitrary initial size Iq. The details are 
given in the Appendix: the result is (l)oo = (Iq + 2)/3, 
which generalizes the continuum result (l)oo = 


Let us recall the zero-temperature coarsening dynam- 
ics of the ID g-state Potts model, starting from a random 
initial configuration. To maximize the speed of the pro- 
gram, we adopt two-sublattice parallel updating. The 
zero-temperature dynamics proceeds as follows: in each 
time step every spin on one of the sublattices changes 
its colour to that of one of its two neighbours with 
equal probability, the two sublattices being updated al- 
ternately. The dynamics of such systems can equivalently 
be formulated in terms of the motions of the domain walls 
as a reaction-diffusion model. The domain walls can be 
considered as random walkers performing independent 
random walks. Whenever two walkers meet, they annihi- 
late (A + A — > 0) with probability l/(q — 1) or aggregate 
(A + A — > A), to become a single walker, with probability 


(q — 2)/(q — 1). The persistence is the probability that a 
fixed point in space has not been traversed by any ran- 
dom walker. For q — 2, the particles only annihilate and 
hence this is equivalent to the Glauber model, whereas 
in the q = oo limit the walkers only aggregate. 

This algorithm is, however, restricted to modelling q- 
state Potts models with q > 2, since q < 2 generates neg- 
ative probabilities, while the general result (|l]) allows any 
real q > 1. In order to simulate values of q in the interval 
(1, 2) we can map the Potts model (as far as persistence 
properties are concerned) on to an Ising representation 
with a fraction 1 j q of the sites initially pointing up, and 
initially persistent, while the down spins are nonpersis- 
tent. Studying the persistence of the minority (majority) 
spins then corresponds to the case q > 2 (q < 2) |17|| . The 
domain walls between the up and down spins behave as 
random walkers, annihilating each other on contact with 
probability one. As far as the persistence is concerned, 
the dynamics of the Potts and Ising models are com- 
pletely equivalent p7| . For q > 2, results obtained from 
both types of simulation are entirely consistent. 

The numerical simulations are performed on a ID lat- 
tice of size 2 x 10 6 with periodic boundary conditions. 
The Ising spins, or Potts variables, occupy the odd sites 
of the lattice, while the random walkers are restricted 
to even-numbered sites. Therefore the effective size of 
the lattice, in terms of Ising/Potts spins, is 10 6 . Each 
walker jumps to an adjacent even-numbered site with 
equal probability, turning any persistent site it hops over 
into a nonpersistent site. The positions of all walkers are 
updated simultaneously at each Monte-Carlo step. The 
initial configurations are chosen to eliminate the possibil- 
ity of any 'crossover' of random walkers when they jump. 
This can be done by placing them only on sites whose 
positions on the lattice are of the form 4fc, where k is an 
integer. The walkers then occupy subsets of the sites 4fc 
and 4fc + 2 alternately. 

For the Potts simulations (q > 2), the lattice is ini- 
tially completely persistent. The random walkers are pe- 
riodically laid down, in clusters of uniform initial size 
Iq. The walkers then execute random walks, aggregating 
or annihilating according to the prescribed probabilities, 
(q — 2)/{q — 1) and l/(q — 1) respectively. The simula- 
tions are performed for initial cluster sizes of 2 and 4. 
The values of q chosen are q = oo, 8 and 4, for which the 
corresponding values of 9 are 8(oo) — 1, 9(8) ~ 0.7942, 
and 6»(4) ~ 0.6315. 

For q < 2 the Ising representation is employed, in 
which persistent sites (up spins) are periodically laid 
down on the lattice, with a fraction l/q of all sites per- 
sistent, and the domain walls execute random walks an- 
nihilating according to A + A — > 0. The value q = 3/2 
was chosen, for which 9 ~ 0.2350. The simulations were 
performed for initial cluster sizes 4 and 8. The Ising sim- 
ulations were also run for q = 4 and 8. The results for 
scaling functions are consistent with the Potts data, but 
the statistics are poorer due to the smaller number of 
clusters generated. For this reason, the data presented 

in Figs. 3-9 were generated from the Potts runs. All the 
simulations (Potts and Ising) were run for 10 000 Monte- 
Carlo steps, and the results averaged over 30 independent 

4000 6000 8000 10000 


FIG. 2. Mean cluster size as a function of time for q = oo 
and initial cluster sizes lo = 2 and 4. 

The simulations investigate the limiting values of the 
average sizes of the persistent clusters, and the distribu- 
tion, n(k, t), of intervals between persistent clusters with 
emphasis on its dynamic scaling form. While the first of 
these lends itself readily to direct measurement, the dy- 
namic scaling form for n(k, t) manifests itself most clearly 
in the data through a study of the (complement of the) 
cumulative distribution. We define 




Inserting the scaling form (Q) and converting the sum to 
an integral in the scaling limit k — > oo, t — > oo with k/t z 
fixed but arbitrary, gives 

I(k,t) = r zz / dk'n(k',t) 



= t~> dxf(x)=t- z g(k/t z ) . (16) 

We now discuss the numerical results for the q = oo 
model. In Figure || we plot the mean cluster size against 
time for initial cluster sizes of 2 and 4. The numer- 
ical results are summarized in Tabic 1. The numer- 
ics clearly support the exact results given in section III 
[and derived in the Appendix as equation)^)], namely 
(0°o = Co + 2)/3 where Iq is the initial cluster size. In 
order to investigate the dynamic scaling form ([l6]) for 
q = oo, we plot, in Figure ^, tl(k,t) against k/t for 
t = 500, 1000, 2000, 4000, 8000, for initial cluster size 
2, where we have used z — 9(oo) = 1 for q = oo. Excel- 
lent data collapse occurs in agreement with the dynamic 
scaling form (|l6|). 







1 qqs _|_ n nnc 


i Q«7 _|_ n 009 
l.OUl In U.UUZ 


i qqq _i_ n nnc 



o m o _i_ n ni o 




1 3945 ± 0007 


2.152 ± 0.001 


2.089 ±0.004 



2.5880 ±0.0002 


4.572 ±0.002 

Table 1 . Average cluster size at late times, (/)oo, for vari- 
ous values of q and various initial cluster sizes lo, where P 
or I indicate whether Potts or Ising representations were 
employed. In all cases the data were averaged over times 
between 1000 and 9000 Monte Carlo steps. 

In section III we argued that the scaling function f(x) 
in (g) should be a simple exponential, which implies that 
g(x) in ( |l6| ) is also a simple exponential. To test this pre- 
diction, we plot ln(tl(k,t)) against k/t in Figure |[ The 
data lie on the expected straight line except at the latest 
times, where the deviations from the line are scatter due 
to statistical noise. Very similar results are obtained for 
initial cluster size 4, confirming that the scaling function 
f{x) is independent of the initial walker density. 

t = 500 

t = 1 000 

t = 2000 

t = 4000 

t = 8000 



FIG. 3. Scaling plot for q = oo, for initial cluster size 
lo = 2. 

In Figure || the mean cluster size is plotted against 
time for various finite values of q and initial cluster size 
l . The numerical results indicate that the average clus- 
ter size tends to a constant (see Table I) in all cases, as 
expected on the basis of the heuristic argument in sec- 
tion II. For q = 8, 9{q) ~ 0.79 > 1/2 and the dominant 
length scale is still determined by the mean distance be- 
tween persistent clusters. We therefore anticipate a scal- 

ing form of the kind given by ( [16]) with z — 9(8) . The 
scaling plot presented in Figure^ shows good data col- 
lapse with this value of z, while Figure indicates that 
g(x), and therefore f(x), is again a simple exponential. 
The data collapse is not as good as for q = oo, which 
we attribute to a poorer separation of the length scales 
L p and L w at the time scales available. This problem 
becomes more acute for q = 4 (see the discussion below) . 





FIG. 4. Same as Figure ^, but presented as a log-linear plot 
to reveal the exponential form of the scaling function. 

.2 4 


q = 3/2, l„ = ' 

q = 3/2, 1 = 
q = 4, l = 4 

q = 8,l = 2 




4000 6000 

FIG. 5. Mean cluster size as a function of time for vari- 
ous values of q and various initial cluster sizes. The Potts 
representation was used for q > 2. 





q = 8, l = 2 

= 500 
= 1000 
= 2000 
= 4000 
= 8000 




FIG. 6. Scaling plot for q — 8 and initial cluster size lo — 2, 
with z = 0(8) ~ 0.7942. 



t = 500 

t- 1000 

t = 2000 

t = 4000 

t = 8000 




FIG. 7. Same as Figure but presented as a log-linear plot 
to reveal the underlying exponential form. 

For q = 4, 6(q) ~ 0.63 > 1/2, so we plot (Figure 
|) t z I(k,t) against k/t z with z = 0(4). This time the 
scaling collapse is definitely not as good as for q = oo 
and q = 8, reflecting an apparent deviation from the dy- 
namic scaling form (16). We believe that this deviation 
reflects the proximity of the length scales L p ~ f - 63 and 
L w ~ t 1 / 2 , which are not sufficiently well separated on 
the timescales achievable in the simulation, and would 
disappear for asymptotically large times. For example, 
L P /L W ~ £0-1315 ^ 3 26 at t = 8 000, while the corre- 
sponding ratio is about 14 for q = 8 and 90 for q — oo. A 
similar effect may explain the apparent deviations from 
universality (e.g. an apparent dependence of z on the 
initial walker density) in the q = 2 simulations of MR. 

In this case = 3/8, and L p /L w ~ i -1 / 8 ~ 0.31 at 
t = 12 000, the largest time reached by the MR simula- 
tions. So in this case also there is not a strong separation 
of length scales at the times achieved in the simulations 
(a point recently emphasized by Manoj and Ray ) . 

t = 500 

t = 1000 

t = 2000 

t = 4000 

t = 8000 




FIG. 8. Same as Figure £j but for q = 
z = 0(4) ~ 0.6315. 


4, with 

-10 - 












FIG. 9. Same as Figure [], but presented as a log-linear plot 
to test the predicted exponential form of g(x). The inset is 
an expanded version of the extreme left of the plot. 

Figure ^| gives the log-linear plot for q = 4. Again, 
the data are consistent with an exponential scaling func- 
tion, except at small scaling variable where there is a 
perceptible deviation from a straight line. In this region, 
however, the data also do not scale at all well either (see 
inset in Figure ||) , which may be another manifestation 
of the poor separation of length scales: the scaling limit 
requires k — ► oo, t — » oo with k/t z fixed. In particular 


the condition k ^> L w ~ t 1 / 2 , which should be satisfied 
for good scaling, is violated at the small values of k/t z 
where the upturn in the data occurs. We will return to 
this point in the Discussion. 

To investigate the spatial distribution of the sites in the 
regime 9 < 1/2 we study q — 3/2, for which 9{q) ~ 0.235. 
In Figure ^| we show that the mean cluster size is a 
constant for initial cluster sizes of 4 and 8. In the 
regime < 1/2 the density of the persistent sites decays 
more rapidly than that of the walkers. The dominant 
length scale is therefore no longer given by the persis- 
tence length, L p ~ t e , but is given instead by the mean 
separation, L w ~ i 1 / 2 , of the the walkers. Hence we ex- 
pect asymptotic dynamic scaling of the form (|2|) and ( |l6| ) 
with z — 1/2. In Figure |l^ we plot t 1 / 2 I{k,t) against 
fc/t 1 / 2 for an initial cluster size la = 4. The numerical 
results give excellent data collapse, supporting the pro- 
posed dynamic scaling form. Apart from a change of 
scale, very similar results are obtained for l = 8, sup- 
porting the universality of the scaling function, i.e. the 
independence of g(x) from the initial walker density. We 
choose q = 3/2 rather than q = 2 in order that the length 
scales L p and L w be reasonably well separated at late 
times: L p /L w ~ f 0265 ~ 0.09 at t = 8 000. 


q = 3/2, l = 4 

= 500 
= 1000 
= 2000 
= 4000 
= 8000 



FIG. 10. Same as Figure U but for q = 3/2, with z = 1/2. 

Figure 10 

reveals a pronounced upturn at small k, 
which can be seen more clearly in the log-linear plot of 
Figure 11 and its inset, suggesting a divergence for k — > 0. 
This is, in fact, expected from the analysis of section II, 
which predicts, for 9 < 1/2, f(x) ~ x~ T for x — ► 0, with 
t = 2(1 — 9), i.e. t is in the range 1 < r < 2. The scaling 
function g(x) is given by g(x) = f dy f(y) ~ x - ^ -1 ' 
for x -> 0, i.e. g(x) ~ jH 1 " 2 ^ = x~ d f , where df = 1-26* 
is the fractal dimension of the persistent set on scales 
smaller than L w [equation (§)]. According to this pre- 

diction, the product 


g(x) should approach a con- 

The small- A: divergence has clearly been removed, the 
function approaching a constant at small x as predicted 
(the erratic behaviour at very small x is due to statistical 


q = 3/2, l = 4 

t = 500 

: icoo 

t = 2000 

t = 4000 

t = 8000 

10 - o.o - 

-0.2 0.0 0.2 0.4 0.6 


FIG. 11. Same as Figure |7l but for 


k/t z 



3/2, and with 

z — 1/2. The inset shows an expanded version of the extreme 
left of the plot, and suggests a singularity for small x = k/t 1 ^ 2 . 

q = 3/2, l = 4 

t = 500 

t = 1000 

t = 2000 

t = 4000 

t = 8000 

stant at small x. This product is shown in Figure hz 

FIG. 12. Same data as Figure [l^, but replotted with or- 
dinate £ 1/2 [fc/t 1/2 ] 1-2e /(M) to show the small-z behavior 
(x — k/t 1 ^ 2 ) more clearly. 

For q — 3/2 all of the numerical simulations have been 
performed using the Ising representation, with periodic 
initial conditions in which clusters of Iq 'up spins' are 
placed at uniform intervals in a 'down spin' background, 
and occupy a fraction 1/q of the sites. For q > 2, the 
Potts representation has mainly been used, but we ex- 
pect all universal properties, such as exponents and scal- 
ing functions, to be the same for the two representations. 
This expectation is confirmed by the results for 9(q) and 


the scaling function g(x). In order to obtain a precise cor- 
respondence between the Ising and the Potts representa- 
tions, however, it is necessary to run the Ising simulations 
with random initial conditions, rather than periodic ini- 
tial conditions which we have mostly employed, and to 
scale the axes appropriately with q to account for the 
different interval sizes in the two cases. This is because 
the reaction-diffusion representation of the Potts model, 
in which walkers annihilate or coalesce with probabilities 
which depend on q, is only an exact representation if the 
Potts states occur in a completely random sequence. This 
is not true for the periodic initial conditions employed in 
the Ising representation, whose correlations spoil the ex- 
act correspondence between Ising and Potts simulations. 
Our expectation, then, is that the Ising representation 
with random and periodic initial conditions should give 
the same scaling functions as the reaction-diffusion imple- 
mentation of the Potts model, but non-universal ampli- 
tudes will be different for the periodic initial conditions. 

_11 i . 1 . 1 1 

10 20 


FIG. 13. Scaling function for q — 8, showing equivalence 
of reaction-diffusion ('Potts') and Ising ('Randomized Ising') 
representations, with z = 0(8), where the axes are qt z I(k,t) 
and k/qt z for the Ising data. The Ising data for periodic ini- 
tial conditions ('Periodic Ising') has the same functional form 
but a different amplitude. 

In order to test the exact equivalence between random 
Ising and Potts simulations we plot, in Figure [l3], the 
results for the q — 8 model using the Potts (reaction- 
diffusion) method. The lattice is initially all persistent 
and the initial cluster size is Iq = 2. This corresponds to 
a sequence of pairs of Potts states in which the members 
of a pair are in same state, but there are no correla- 
tions between pairs (except that neighboring pairs must 
be in different states). The random initial conditions for 
the Ising simulation are generated from the Potts initial 
state by setting all pairs in Potts state T' (say) to be 
up spins and all other pairs to be down spins. In the 
Potts simulations we keep track of all persistent sites, 

while in the Ising simulation we track the persistence of 
the spins initially 'up'. The interval sizes (between per- 
sistent clusters) are then naturally larger by a factor q 
in the Ising simulations relative to the Potts simulations. 
Figure [H?] shows both scaling functions in log- linear form, 
for t = 10 4 , with this factor of q scaled out (i.e. qt z I(k, t) 
is plotted against k/qt z for these data). The data over- 
lap almost perfectly. The Ising data for a periodic initial 
condition are consistent with the same scaling function 
(exponential), but with a different amplitude. 


In this paper we have investigated the nature of per- 
sistent structures in the coarsening dynamics of ID Potts 
models. A central concept has been the existence of 
two characteristic length scales, the mean separation, 
L w ~ t 1 / 2 , of domain walls (or 'walkers') and the 'persis- 
tence length', L p ~ t e , which measures the mean separa- 
tion of persistent sites or clusters. The focus of our at- 
tention has been the distribution function, n(k, t), for the 
number of intervals (between clusters of persistent sites) 
of length k. This distribution has the scaling form (j^), 
with characteristic length scale, t z , which is the larger of 
L w and L p , i.e. z = max(l/2,0). Within the ID Potts 
model both regimes, 9 < 1/2 and 9 > 1/2, can be ac- 
cessed by varying the number of Potts states, q. From 
the general result, (|IJ), we see that 9(q) is a monotoni- 
cally increasing function with, q c = 2.70528 . . . marking 
the boundary between the two regimes. 

The regime 9 > 1/2 is conceptually simpler. The 
q — oo limit can been solved exactly (section III), and 
results for the long-time limit of the mean cluster size ob- 
tained. The fact that the locations of the surviving clus- 
ters are statistically independent in the scaling regime 
leads to the result that the scaling function f(x) in (||) is 
a simple exponential. We have argued that the same re- 
sult should hold for all 9 > 1/2, since in this regime there 
are, on average, many walkers between each pair of per- 
sistent clusters. The relevant walkers for this argument 
(those which are turning persistent sites in neighboring 
clusters nonpersistent) are uncorrelated since their sep- 
arations, of order L p , are much larger than the typical 
separation, L w , of neighboring walkers. On this basis we 
expect the asymptotic scaling function f(x) to be expo- 
nential for all 9 > 1/2 (i.e. q > q c ). The data for q = 8 
(Figure |7|) and q = 4 (FigureS) are consistent with an ex- 
ponential form for g(x) = J f(x), although the scaling 
is not perfect and there is a small upturn in the scaling 
function at small scaling variable for q = 4. We attribute 
these features to an imperfect separation of length scales 
on the time scales achieved in the simulations, and con- 
jecture that the true scaling function is exponential for 
all 9 > 1/2. 

The case 9 < 1/2 is more tricky. In this case the 
persistent clusters outnumber the walkers. The scaling 


function f(x) is clearly not a simple exponential, though 
it seems (Figure [ll]) to have an exponential tail. There 
is a small- argument singularity of the form f(x) ~ x~ T , 
with t = 2(1 — 9). This is related to the fractal dimen- 
sion, df = 1 — 29, of the persistent sites: df = r— 1. Note 
that the borderline, 9 = 1/2, between the two regimes oc- 
curs at df = 0. The existence of a small- :r singularity for 
9 < 1/2 raises the possibility of an alternative scenario for 
9 > 1/2, in which the x~ T singularity, with r = 2(1 — 9), 
persists for 9 > 1/2, where it becomes an integrable sin- 
gularity. The small- a; singularity in g(x) would then take 
the form of a cusp: g{x) = g(0) — Ax 26 ^ 1 + ••-. We have 
not been able to rule out this scenario from the data, but 
think it unlikely for the reasons given elsewhere in this 
paper. Further insight could be obtained if it were pos- 
sible to perform an expansion around the soluble q = oo 
limit to first order in 1 /q, but we leave this as a challenge 
for the future. 

We conclude by discussing briefly the possibility of the 
existence of these two qualitatively different regimes in 
spatial dimension d > 2. First we generalize the re- 
sult (|8|) , relating df and 9, to any dimension. Starting 
from (Q), the result F(x) ~ x~ 2e follows generally, and 
the number of persistent sites within a distance R of a 
given site is estimated as §^ r d ~ 1 dr r~ 29 ~ R d f , where 
df = d — 29. Generalizing still further, if the coarsening 
exponent is <p, rather than 1/2, the result is 

d f =d- 


(the d = 2 version of this result is given in jli| , based on 
the same reasoning). Clearly this result requires 9 < dxf>, 
since df cannot be negative. If this inequality is violated, 
as in the ID Potts model with 9 > 1/2, the persistent 
sites no longer have a fractal structure but become point- 
like objects, with mean density ~ t~ e <C t~ d ^ ~ L~ d , 
where L c ~ t<P is the coarsening length scale. Cases 
where 9 < dtp are easy to find, e.g. in the coarsening 
of the 2D Ising model Q , the 2D diffusion equation || , 
or the time-dependent Ginzburg-Landau equation in 2D. 
These all exhibit fractal persistent structures with the ex- 
pected fractal dimension ]il||l8| |. It would be interesting 
to look for examples, in addition to the ID Potts model, 
where one can have 9 > dip. 


This work was supported by EPSRC (UK). 













B. Derrida, A. J. Bray, and C. Godreche, J. Phys. A 27, 
L357 (1994). 

B. Derrida, V. Hakim, and V. Pasquier, Phys. Rev. Lett. 

75, 751 (1995); J. Stat. Phys. 85, 763 (1996). 

D. Stauffer, J. Phys. A 27, 5029 (1994). 

S. N. Majumdar, C. Sire, A. J. Bray, and S. J. Cornell, 

Phys. Rev. Lett. 77, 2867 (1996); B. Derrida, V. Hakim, 

and R. Zeitak, ibid. 77, 2871 (1996). 

J. Krug, H. Kallabis, S. N. Majumdar, S. J. Cornell, A. 

J. Bray, and C. Sire, Phys. Rev. E 56, 2702 (1997). 

A. J. Bray, B. Derrida, and C. Godreche, Europhys. Lett. 
27, 175 (1994); P. L. Krapivsky, E. Ben-Nairn, and S. 
Redner, Phys. Rev. E 50, 2474 (1994); S. N. Majumdar 
and C. Sire, Phys. Rev. Lett. 77, 1420 (1996). 

J. Cardy, J. Phys. A 28, L19 (1995); E. Ben-Nairn, Phys. 
Rev. E 53, 1566 (1996); M. Howard, J. Phys. A 29, 3437 
(1996); S. J. Cornell and A. J. Bray, Phys. Rev. E 54, 
1153 (1996). 

M. Marcos-Martin, D. Beysens, J-P. Bouchaud, C. 
Godrche, and I. Yekutieli, Physica (Amsterdam) 214D, 
396 (1995). 

W. Y. Tarn, R. Zeitak, K. Y. Szeto, and J. Stavans, Phys. 
Rev. Lett. 78, 1588 (1997). 

B. Yurke, A. N. Pargellis, S. N. Majumdar, and C. Sire, 
Phys. Rev. E 56, R40 (1997). 

S. N. Majumdar, A. J. Bray, S. J. Cornell, and C. Sire, 
Phys. Rev. Lett. 77, 3704 (1996); K. Oerding, S. J. Cor- 
nell, and A. J. Bray, Phys. Rev. E 56 R25 (1997). 
G. Manoj and P. Ray, preprint, cond-mat/9901130 

(1999); preprint, |zond-mat/0003203| (2000) 

D. H. Zanette, Phys. Rev. E 55, 2462 (1997). 

G. Manoj and P. Ray, J. Phys. A 33, L109 (2000). 

S. Jain and H. Flynn, preprint, cond-mat/0004148 


E. Ben-Nairn, L. Frachebourg, and P. L. Krapivsky, Phys. 
Rev. E 53, 3078 (1996). 

S. J. Cornell and A. J. Bray, unpublish ed. 

See for 

some illustrations of the persistent structures which arise. 


In this Appendix we compute the probability that a 
cluster of initial size Iq survives to time t, and the mean 
size of surviving clusters, for the process A + A — > A. 

First note that initially there is a random walker at 
each end of the domain. At each time step the walkers 
move independently left or right with probability 1/2, so 
we can treat each walker independently. Consider, there- 
fore, a single walker moving at discrete time steps on a 
discrete ID lattice, starting, at time t = 0, at position 
r. Let the 'origin' be the point r = 0, and let P r {t) be 
the probability that the walker has not yet reached the 
origin at time t. Clearly, 


[1] For a recent review see S. N. Majumdar, Curr. Sci. India 
77, 370 (1999). 

Pi(t) = -P 2 (t-1) 
Pr{t) = \[P r -i{t-l) 

P r+1 (i-l)], r>2. (18) 


We are interested only in the limit t — > oo. In this limit, 
we know from standard random walk theory that every 
P r (t) decays like i -1 / 2 , with an r-dependent amplitude. 
To leading order in i -1 / 2 , therefore, the t-dependence 
drops out of equations (p"8|), which then become equa- 
tions for the amplitudes. By inspection, the solution in 
this regime is 

P r (t)=rP 1 (t) 


Now consider a walker starting immediately to the 
right of a cluster of persistent sites. The probability 
that after t steps the walker has jumped over exactly r of 
these (making them nonpersistent) is P r +\{t) — P r {t) = 
P\ (t) , where the final result follows from ([H]) , to leading 
order for large t. The same result holds for a walker start- 
ing immediately to the left of the cluster. The probability 
that I sites remain persistent after time t is Pi(t) 2 times 
the number of ways of partitioning the cluster of length Iq 

into 3 sections, with the central section of length I (and 
zero lengths are allowed for the outer sections). This 
number is Iq — I + 1. So the probability of the cluster 
surviving and having length I is 

p lo (l,t) = (l -l + l)Pi(t) 2 , 

a generalization of (|l2] ) to the discrete system. 
The survival probability of the cluster is 

while the mean cluster size is 

(l) _ TtitPh(i,t) _ Go + 2) 


lo ^ 
p s urv{t)=J2Plo( l ^) = ^o(lo + l)Pl(tf , (21)